Abstract:
Memes are highly prevalent in online social media content, making them a vital source for analysing their power in diffusing ideas online. Memes combine the appeal of visuals with compelling messages, which have the potential to be highly influential. Previous work on memes has focused on meme analysis in terms of harmfulness, hate, and offence and looking at the entities targeted by memes as heroes, villains and victims. We look for a broader understanding of memes in different contexts. Through the course of this project, we looked at four main problems to analyse memes: 1. Generating explanations for semantic role labels of entities in a meme, 2. Intent Classification to study intent behind meme dissemination patterns, and 3. To formulate a Multimodal-Question Answering setup to understand who is targeted in a meme and with a particular intent in the dissemination, and 4. Identifying fine-grained depressive symptoms in memes. To this end, we have curated datasets that include labels for semantic role labels (Hero, Villain Victim), intent of distribution (Persuasive, Grassroots, Public Discourse) and natural language explanations for these connotations. In addition, we have contributed to existing datasets on depressive symptoms in memes by cleaning them and making them more useful. Through a combination of these problems, we aim to get a broader understanding of memetic content in general. This is a step towards a future where a generalised model may be as good as specialised models for each individual task we have defined. With meme content, the biggest challenge has always been the capacity to inform about the irony, humour and implied meanings present in the memes to the model and through this work, we aim to explore some task specific methods of working on this problem.